Add pass@1 for GPQA-D and MATH-500 #698
Merged
+113
−13
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
lewtun
commented
May 1, 2025
| max_model_length: PositiveInt | None = None # maximum length of the model, ussually infered automatically. reduce this if you encouter OOM issues, 4096 is usually enough | ||
| swap_space: PositiveInt = 4 # CPU swap space size (GiB) per GPU. | ||
| seed: PositiveInt = 1234 | ||
| seed: NonNegativeInt = 1234 |
There was a problem hiding this comment.
Needed to allow seed=0 in the model args
lewtun
commented
May 1, 2025
| few_shots_select=None, | ||
| generation_size=32768, | ||
| metric=[ | ||
| Metrics.expr_gold_metric, |
There was a problem hiding this comment.
This was causing redundant computation since we can get the same result from pass@1 (n=1)
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
lewtun
changed the title
lewtun
commented
May 1, 2025
|
|
||
| def compute(self, golds: list[str], predictions: list[str], **kwargs) -> dict[str, float]: | ||
| def compute( | ||
| self, golds: list[str], predictions: list[str], formatted_doc: Doc = None, **kwargs |
There was a problem hiding this comment.
I had to pass formatted_doc in order to enable pass@1 with GPQA
NathanHB
approved these changes
May 5, 2025
1 task
hynky1999
pushed a commit
that referenced
this pull request
May 22, 2025
* Add pass@1 for GPQA-D and clean up AIME * Add pass@1 for math_500 * Add pass@1 for MATH-500 * Update test * Fix
NathanHB
pushed a commit
that referenced
this pull request
Sep 19, 2025
* Add pass@1 for GPQA-D and clean up AIME * Add pass@1 for math_500 * Add pass@1 for MATH-500 * Update test * Fix
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
For GPQA Diamond and MATH-500, it is customary to report
pass@1instead of raw accuracy. In this PR, I've:nwas chosen to have ~2k total completions to mitigate variance. For a 7B model this takes about 15' on 1 node with DP=8, which seems like a reasonable tradeoff in variance vs computen=64. I realise I down-scaled this originally ton=32, but after reading the Phi-4-reasoning tech report I think it's better to increase to mitigate variance as much as possible since AIME24 is often used for model selection during training. (See plot below where they show the pass@1 variance can be rather large across repeated runs if T is medium/high as is common for LRMs)For GPQA, I've checked that my pass@1 (n=1) metric is equivalent to the old accuracy based one (as expected) for
Qwe2.5-7B-Instruct:Please let me know if I should updated any other benchmarks to align with these changes!